Research On Key Algorithms Of Dynamic Gesture Recognition

Because of many advantages such as flexibility,convenience and rich information,gesture has been widely used in Human-Computer Interaction,in the fields of somatosensory game and live broadcasting,the application of gesture recognition ushers in the new opportunities.However,some problems still restricts recognition accuracy,which include complex background,gesture overlapping and occlusion.According to the above discussion,in the research the traditional frame difference method and skin color detection method are improved,gesture feature extraction method based geometric and track feature is refined,and the gesture recognition accuracy is increased under complex background.The research mainly includes three parts,which includes hand area detection and segmentation,gesture feature extraction and real-time recognition based on Support Vector Machine(SVM):(1)Hand accurate detection and segmentation based on the fusion of frame difference method and skin color detection method.First,the traditional three-frame method is improved through adding illumination suppression coefficient,and the influence of illumination is reduced in moving object detection under complex environment;Then,an adaptive threshold selection algorithm is proposed,and skin area detection is achieved in YCbCr space.Through the fusion of the above two methods,the problem of detecting incomplete hand shape,which is caused by some factors such as hand occlusion and deformation in the process of movement,has been solved well.(2)Multiple gesture features fusion.Firstly,Fourier descriptor,normalized rotational inertia and invariant moment are used as the geometric feature to distinguish different hand shapes;Secondly,the variation of hand spatial position can be used as gesture feature,different types of gestures are identified according to moving trail direction angle variation;Finally,the final gesture feature can be acquired through fusing the extracted geometric feature and trajectory feature.(3)Gesture recognition based on SVM.The used SVM selects Radial Basis Function(RBF)as kernel function,which is applicable to the samples of different scales and dimensions,and has wide convergence domain,in addition,the tree classifier is used to classify in the SVM.Through selecting the extracted gesture feature vectors from sample database to train the designed system,the classifier model can be acquired,and the recognition of different types of gestures is achieved.In experiment process,different gestures of 100 participants are collected under different backgrounds,then,training and testing are achieved based on the proposed method,under complex background,the recognition rates of all different gestures can exceed 85%,on the contrary,the average recognition rate can reach 96% under simple background,Experiment results show that the proposed gesture recognition algorithm can acquire good effect.The proposed dynamic gesture recognition algorithm can be applied in many fields,such as deaf-mutes education,remote control and military modern sign language interaction.In addition,the designed moving area and skin area detection methods have good reference value for solving other pattern recognition problems,such as pedestrian detection,face recognition,etc.